Datacenters are a critical component of the modern internet, responsible for processing and storing tremendous amounts of data in the “cloud.” Datacenters also provide the computational power needed for handling “big data,” a growing segment of the U.S. economy. Currently, datacenters consume more than 2.5% of U.S. electricity and this figure is projected to double in about eight years due to the expected growth in data traffic. There are many approaches to improving the energy efficiency of datacenters, but these strategies will be limited by the efficiency with which information travels along metal interconnects within the devices in the datacenter—all the way down to the computer chips that process information. Unlike metal interconnects, photonic interconnects do not rely on electrons flowing through metal to transmit information. Instead, these devices send and receive information in the form of photons—light—enabling far greater speed and bandwidth at much lower energy and cost per bit of data. The integration of photonic interconnects will enable new network architectures and photonic network topologies that hold the potential to double overall datacenter efficiency over the next decade.
Project Innovation + Advantages:
The University of California, Berkeley (UC Berkeley) will develop a new datacenter network topology that will leverage the energy efficiency and bandwidth density through the integration of silicon photonics into micro electro-mechanical system (MEMS) switches. Today's datacenter architectures use server nodes (with processor and memory) connected via a hierarchical network. In order to access a remote memory in these architectures, a processor must access the network to get to a particular server node, gaining access to the local memory of that server. This requires the remote server processor to be awake at all times in order to service the remote request. The processor-to-memory network has many stages and long latency, which results in significant energy waste in processor and memory idling on both sides of the network. The IceNet network is designed to achieve ultra-low latency connectivity between processor nodes and memory, drastically reducing energy wasted during system idling. A key component to the team's design is their LightSpark active laser power-management system. In addition to guiding the laser power where it is needed, the LightSpark module enables both wavelength and laser redundancy, increasing the robustness of the system. In total, the IceNet network will enable dramatic improvements in datacenter system efficiency, allowing for fine-grain power control of processors, links, and memory and storage components.
If successful, developments from ENLITENED projects will result in an overall doubling in datacenter energy efficiency in the next decade through deployment of new photonic network topologies.
The United States is home to much of the world’s datacenter infrastructure. Photonic networks add resilience that can bolster the energy security of this critical driver of economic activity.
Reducing the overall energy consumption of datacenters cuts energy-related emissions per bit of data transmitted and processed.
Photonic networks can lower the costs associated with operating datacenters, improving American economic competitiveness in this fast-developing area.